Selectel
AI news source. Articles are auto-selected and adapted by Hamidun News editors.
Latest publications

Pinterest replaces its profiler patchwork with Intel’s open-source gProfiler
When services written in Java, Python, C++, Go and Node.js run on hundreds of thousands of machines, standalone tools like perf and py-spy are no longer enough. Pinterest has opted for a single open-source profiler.

Tencent trains Hy4 to control computers through two interfaces: GUI and CLI
Hy4 can understand images and work with both graphical interfaces and the command line. For developers, that makes it worth testing whether the agent can handle an entire task, from browser to terminal.

GPT-6 Cyber Finds 81 New Vulnerabilities in Firefox and Prepares Fixes
AI can now check code for vulnerabilities and back up its findings with reproducible tests. OpenAI has shown GPT-6 Cyber's results on Firefox and Linux, but access remains limited for now.

Google Cuts Gemini Omni Video Generation Price by 30% — to $0.07 per Second
A fifteen-second 720p clip now costs $1.05 per generation. For video services it's a reason to recalculate costs, but the SLA promise in the headline deserves separate verification.

OpenAI makes Git pushes nearly five times faster with Pando
When agents write code in parallel, the Git server has to accept more changes and deliver them to test environments. OpenAI rebuilt its backend while keeping familiar Git commands intact.

Moonshot releases Kimi K3.1 Pro weights with a 1 million-token context window
Kimi K3.1 Pro can now be deployed on your own infrastructure to process code alongside screenshots. But its 2.8 trillion parameters call for careful estimates of memory requirements and costs.

The best AI model completes just 12% of work tasks in the new GDPx benchmark
SemiAnalysis has released GDPx, a test of agents on tasks businesses actually pay for. Even the best performer, Claude Opus 5.5, completed just 12% of tasks.

GPT-6 Astra Fixed Qwen's Weights and Lifted Its MATH-500 Score to 48.1%
A small Qwen model kept getting stuck in repetition instead of answering. MIT researchers had Astra fix its weights: the score on a math benchmark rose from 7.9% to 48.1%. But there is a caveat.

Axion 1.0-35B scores 52.2% on HLE as its creators release agent training code
An agent needs to find facts, verify sources and perform calculations. Axion offers an open recipe for that training, but its 52.2% HLE score does not yet prove the model can replace your API.

Revolut moves 2000 engineers to self-hosted GLM and cuts AI coding costs by 90%
The British fintech dropped Claude Code and Codex, deployed the open GLM model on its own GPUs, and reports costs of around $0.30 per million tokens. The figures have yet to be independently verified.

Engineer Compares 140 W and 240 W GaN Chargers, Then Starts Building His Own PD Module
A 240 W GaN charger from a marketplace costs 15 times less than a proven 140 W one. A Habr author checks what the label number really means and builds their own Power Delivery module.

Selectel Walks Through Installing Jenkins on a Cloud Server: From Resources to a Pipeline
A Selectel engineer shows step by step how to order a virtual server, install Jenkins, and build a first pipeline that automatically builds an application from Git.

Selectel Tests Mac mini M4 Pro as a Local LLM Server for OpenClaw
Selectel tested a Mac mini with the M4 Pro chip as a portable LLM server running the OpenClaw agent. Apple Silicon's unified memory removes the main barrier to local inference: the lack of VRAM.

Benchmarks Lie: Why Businesses Are Moving to Open Weights and In-House Inference
Enterprises are losing faith in third-party APIs: regulators, data leaks and unreliable benchmarks are pushing companies toward open models and their own infrastructure.